Report: Meta's Llama 3.1 Memorized 42% of the first Harry Potter book
The discovery has been published in a joint research paper
2 min. read
Published on
Read our disclosure page to find out how can you help Windows Report sustain the editorial team. Read more

If you thought AI was just generating fresh content, think again. Meta’s latest model, Llama 3.1, has been quietly memorizing and repeating copyrighted books, and not just a few lines.
A new academic research paper from Stanford, Cornell, and West Virginia University reveals that Meta’s Llama 3.1 can spit out nearly half of Harry Potter and the Sorcerer’s Stone in 50-word segments. That’s 42% of the book, to be precise.
Do note that the earlier versions of the Llama model didn’t go this far. Llama 1 could only recall about 4% of the same book. But Meta’s Llama 3.1 doesn’t just stop at Hogwarts. It also pulls from titles like 1984 and The Hobbit, all part of a dataset called Books3—a collection of thousands of copyrighted works.
Smaller authors may be less at risk. The AI remembered only 0.13% of Sandman Slim, a lesser-known novel. That kind of uneven retention could make future lawsuits messier. The legal argument now hinges on which books are being copied—and how exactly the AI is reproducing them.
Also read: Meta wants AI to fully run its ads by 2026
So, how did this happen? Researchers suspect the books were overused in training or scraped from fan sites and reviews. Some think minor training tweaks might have caused the AI to memorize more than it should.
User forum
0 messages